DexRelay · March 2026

The future of building is on your phone.

DexRelay turned my iPhone into the control plane for Codex on my Mac. I code from my phone all day now, in transit, at coffee shops, wherever. The Mac is the machine. The phone is the company.

25 Active projects
24/7 Operating time
1 Surface
01 — What Changed
02 / 07

One phone. Everything running.

Prompts, approvals, project context, build status, install links, all in one surface. Idle time became operating time. This isn't mobile coding. It's mobile orchestration of a real runtime.

Remote Codex threads

Connect to any Codex thread on the Mac over WebSocket. Approve commands, browse project files, queue follow-ups, without touching the desk.

Build & install iOS apps

Trigger Xcode builds, publish OTA install artifacts, and open them on the phone. Real compilation, real binaries, no simulator.

OTA over Tailscale

Phone and Mac on the same tailnet. Works from anywhere, home, transit, hotel wifi. Private, zero-config networking.

Speech to code

Dictation as the prompt interface. Talk direction into the phone, the Mac executes. Talking to code is the obvious next interface.

02 — Architecture
03 / 07

iPhone control plane.
Mac execution host.

Three layers. The phone gives direction. The bridge normalizes. The Mac executes.

Layer 1

iPhone App

SwiftUI client. Lists threads, starts turns, streams activity, handles approvals, attachments, dictation, and build actions.

→ connects to :4600
Layer 2

DexRelay Bridge

Node bridge that normalizes WebSocket behavior for iOS, trims oversized history, exposes helpers for shell execution and media upload.

local/exec · local/uploadMedia
Layer 3

Codex App-Server

Mac runs thread and turn logic against local source trees, tools, files, and git state. This is where execution happens.

app-server on :4500
03 — The Loop
04 / 07

DexRelay was built using DexRelay.

The phone client steers Codex. Codex edits the phone client. I approve, build, install, use the improved app, then use it to build the next improvement. The tool compounds itself.

Why that's 10x

The delay between idea → prompt → approval → build → install → reaction collapses. Instead of losing momentum every time I leave the desk, the system keeps compounding through the day. A product becomes serious when it can build more of itself.

It's not that I type on glass. It's that I keep the factory running from the phone.
DexRelay operating model · 05 / 07
04 — Interfaces Shifting
06 / 07

Terminal. IDE. PRD.
All demoted.

Terminal

→ Service layer

The shell still exists but stops being the primary interface. It becomes infra the agent calls.

IDE

→ Execution context

Xcode still matters, but as a toolchain, not the center of attention. The control layer moved upward.

PRD

→ Live steering

Instead of freezing intent in a static doc, you keep threads, plans, and approvals alive. The spec updates while the product is being built.

Typing

→ Orchestrating

The scarce skill is no longer "can you code?" It's: can you steer intelligence toward something people actually want?

05 — Org in Threads · 06 — Declaration
07 / 07

Multi-nested threads mimic an org chart.

You don't need a giant org to create functional specialization. You need a clean thread hierarchy, reusable skills, clear context, and an operator who can assign work well.

You — The Operator

Mission, constraints, release target

Product Thread

Plan, user value, scope

Design Thread

Interface, hierarchy, taste

IC Subthreads

Impl, QA, release, content

The future builder looks less like a solo coder and more like a CEO managing a bench of VP agents.

"Can you build a product?" is rapidly becoming table stakes. What matters now: distribution, product instinct, retention, and taste.

The first limiter is token budget. The second is willingness. The third is taste. Intelligence without taste just produces more output.

Soon nobody will care that someone can build with AI. The question will be whether people use what they built.

DexRelay · March 2026

The future of building is on your phone.

DexRelay turned my iPhone into the control plane for Codex on my Mac. I code from my phone all day now, in transit, at coffee shops, wherever. The Mac is the machine. The phone is the company.

25Active projects
24/7Operating time
1Surface
01 — What Changed

One phone. Everything running.

Prompts, approvals, project context, build status, install links, all in one surface. Idle time became operating time. This isn't mobile coding. It's mobile orchestration of a real runtime.

Remote Codex threads

Connect to any Codex thread on the Mac over WebSocket. Approve commands, browse project files, queue follow-ups, without touching the desk.

Build & install iOS apps

Trigger Xcode builds, publish OTA install artifacts, and open them on the phone. Real compilation, real binaries, no simulator.

OTA over Tailscale

Phone and Mac on the same tailnet. Works from anywhere, home, transit, hotel wifi. Private, zero-config networking.

Speech to code

Dictation as the prompt interface. Talk direction into the phone, the Mac executes. Talking to code is the obvious next interface.

02 — Architecture

iPhone control plane.
Mac execution host.

Three layers. The phone gives direction. The bridge normalizes. The Mac executes.

Layer 1

iPhone App

SwiftUI client. Lists threads, starts turns, streams activity, handles approvals, attachments, dictation, and build actions.

→ connects to :4600
Layer 2

DexRelay Bridge

Node bridge that normalizes WebSocket behavior for iOS, trims oversized history, exposes helpers for shell execution and media upload.

local/exec · local/uploadMedia
Layer 3

Codex App-Server

Mac runs thread and turn logic against local source trees, tools, files, and git state. This is where execution happens.

app-server on :4500
03 — The Loop

DexRelay was built using DexRelay.

The phone client steers Codex. Codex edits the phone client. I approve, build, install, use the improved app, then use it to build the next improvement. The tool compounds itself.

Why that's 10x

The delay between idea → prompt → approval → build → install → reaction collapses. Instead of losing momentum every time I leave the desk, the system keeps compounding through the day. A product becomes serious when it can build more of itself.

It's not that I type on glass. It's that I keep the factory running from the phone.
DexRelay operating model
04 — Interfaces Shifting

Terminal. IDE. PRD.
All demoted.

Terminal

→ Service layer

The shell still exists but stops being the primary interface. It becomes infra the agent calls.

IDE

→ Execution context

Xcode still matters, but as a toolchain, not the center of attention. The control layer moved upward.

PRD

→ Live steering

Instead of freezing intent in a static doc, you keep threads, plans, and approvals alive. The spec updates while the product is being built.

Typing

→ Orchestrating

The scarce skill is no longer "can you code?" It's: can you steer intelligence toward something people actually want?

05 — Org in Threads

Multi-nested threads mimic an org chart.

You don't need a giant org to create functional specialization. You need a clean thread hierarchy, reusable skills, clear context, and an operator who can assign work well.

You — The Operator

Mission, constraints, release target

Product Thread

Plan, user value, scope

Design Thread

Interface, hierarchy, taste

IC Subthreads

Impl, QA, release, content

The future builder looks less like a solo coder and more like a CEO managing a bench of VP agents.

06 — Declaration

Building becomes cheap.
Adoption becomes the test.

"Can you build a product?" is rapidly becoming table stakes. What matters now: distribution, product instinct, retention, and taste.

The first limiter is token budget. The second is willingness. The third is taste. Intelligence without taste just produces more output.

Soon nobody will care that someone can build with AI. The question will be whether people use what they built.